Isn't it just learning to map specific words, from the "knife world", to the correct class? If so, a simple dictionary would fit. What I think is a better way to validate is to split train/validation by words used presented in NER classes (like, it should be able to find new brands never seen before). It is a interesting problem.

That's a key piece of the article. He 'trusts' Gemini to classify the posts, and never hand validates anything.

And he continues building on top of that shakey trust. Is this a good idea?

It just depends on what you're doing: sounds like he's just having fun with a hobby, so it's harmless.

All he's really done is work more efficiently to save token cost and time. But he hasn't validated anything, so there's no telling if he's wasting time or not. It's a hobby though....

It wasn’t as simple as just mapping brands… for one I don’t have a complete dictionary of all the brands that could be mention.

Other problems are:

Brand names that match common English words, brand names that are also item models used by another brand… etc

If you are planing to watch the eclipse, a good resource is to look at http://xjubier.free.fr/en/site_pages/SolarEclipsesGoogleEart... In 2028 it will go straight to Sydney, Australia

Nowadays we use LLMs mostly for doing agentic-based work. LLMs new Pareto frontier only make the headlines if they push the boundaries on benchmarks that are deterministic tasks. So models are encouraged to focus on these deterministic tasks that are, in nature, structured texts. I think that this makes models more “plastic” or “polished”, as opposed to natural and pleasant to read. User-based benchmarks, like LLM Arena, are for me the best we can do in order to rank models in this way, but come with its own drawback (subjective evaluation, prone to spam or techniques to promote a giving model).

It is good to se big companies like Microsoft launching LLMs. They have large amount of compute power and good scientists to create useful models.

Microsoft has been releasing LLMs for years.

Sort of. Phi models were just trained on GPT outputs though.

For those that don't know about this. Phi was announced with a paper called "Textbooks are all you need". What they did was use GPT 3.5 and created synthetic textbook chapters and exercises.

They also did some more interesting work like showing very small models can be coherent as long as you have very simple children's book style training data (TinyStories is pretty famous).

Lots of these ideas are still used. Learning facts at scale with active reading is an ICLR 2026 paper from Meta AI that does a lot of similar work.

By design. The whole point of Phi is the "textbooks is all you need" theory on curated training data, as opposed to kitchen sinks.

They were mostly distilled or fine-tuned OAI models.

huh? The granite series isn't distilled

Granite is IBM

Ah snap. You’re right of course

And occasionally un-releasing them like with WizardLM.

Any plans to port to sglang or vLLM?

vllm-omni support is on the way : )

It seems to be something related the moving average calculation. So it is just a glitch on the chart.

This guy seems to be talking seriously.

Not to demerit the recording, but I felt more nostalgic for the last sentence of the article "Sometimes, the internet is good" than for the musics itself.

We all know it... but I think they were very bold in this warning about using your private messages to train public models. _Your messages with AIs will be used to improve AI at Meta. Don't share information, including sensitive topics, about others or yourself that you don't want the AI to retain and use_

meta doesn't exactly instill confidence on using personal data responsibly. hard pass